WO2008106615A1 - Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques - Google Patents

Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques Download PDF

Info

Publication number
WO2008106615A1
WO2008106615A1 PCT/US2008/055352 US2008055352W WO2008106615A1 WO 2008106615 A1 WO2008106615 A1 WO 2008106615A1 US 2008055352 W US2008055352 W US 2008055352W WO 2008106615 A1 WO2008106615 A1 WO 2008106615A1
Authority
WO
WIPO (PCT)
Prior art keywords
spatial
temporal
node
program code
patterns
Prior art date
Application number
PCT/US2008/055352
Other languages
English (en)
Inventor
Robert G. Jaros
Jeffrey L. Edwards
Dileep George
Jeffrey C. Hawkins
Original Assignee
Numenta, Inc.
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Numenta, Inc. filed Critical Numenta, Inc.
Publication of WO2008106615A1 publication Critical patent/WO2008106615A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs

Definitions

  • the present invention is directed toward the field of machine learning using
  • HTM Hierarchical Temporal Memory
  • Machine learning has generally been thought of and attempted to be implemented in the context of artificial intelligence.
  • Artificial intelligence depends on algorithmic solutions (e.g., a computer program) to replicate particular human acts and/or behaviors.
  • algorithmic solutions e.g., a computer program
  • the study of neural networks is a sub-area of artificial intelligence which attempts to mimic certain human brain behavior by using individual processing elements that are interconnected by adjustable connections.
  • the present invention provides a computer-implemented hierarchical network comprising a plurality of spatio-temporal learning nodes.
  • Each spatio-temporal learning node in the computer implemented hierarchical network comprises a spatial pooler and a temporal pooler.
  • the spatial pooler is adapted to receive a sensed input pattern.
  • the spatial pooler is further adapted to generate a first set of spatial probabilities associated with a set of spatial cooccurrence patterns, wherein each spatial co-occurrence pattern represents a first set of one or more sensed input patterns and each spatial probability in the first set of spatial probabilities indicates the likelihood that the sensed input pattern has the same cause as a spatial cooccurrence pattern.
  • the temporal pooler is adapted to receive the first set of spatial probabilities from the spatial pooler.
  • the temporal pooler is further adapted to generate a set of temporal probabilities associated with a set of temporal groups based at least in part the first set of spatial probabilities, wherein each temporal group comprises one or more temporally co-occurring input patterns and each temporal probability indicates the likelihood that the sensed input pattern has the same cause as the one or more temporally co-occurring input patterns in a temporal group.
  • the temporal pooler is further adapted to transmit the set of temporal probabilities to a parent node in the hierarchical network of nodes.
  • the present invention provides a computer-readable storage medium encoded with computer program code for the computer-implemented hierarchical network comprising a plurality of spatio-temporal learning nodes described above.
  • the present invention provides a computer-implemented hierarchical network comprising a plurality of spatio- temporal learning nodes assigned to a plurality of hierarchical levels. Each spatio-temporal learning node is adapted to receive a plurality of input patterns. Each spatio-temporal learning node is further adapted to execute a temporal pooling algorithm to identify temporal groups, wherein each temporal group comprises one or more input patterns which temporally co-occur.
  • Each spatio-temporal learning node is further adapted to execute a spatial pooling algorithm to identify subsets of one or more input patterns that can be represented using a spatial co-occurrence pattern, wherein the spatial pooling algorithm executed by each node is defined by a hierarchical level the spatio-temporal learning node is assigned to.
  • the present invention provides a computer-readable storage medium encoded with computer program code for the computer-implemented hierarchical network comprising a plurality of spatio-temporal learning nodes assigned to a plurality of hierarchical levels described above.
  • Figure 1 shows a flow of data between an object and a human.
  • FIG. 2 shows a HTM in accordance with an embodiment of the present invention.
  • FIG. 3 shows a spatio-temporal learning (STL) node in accordance with an embodiment of the present invention.
  • Figure 4 shows a STL top node in accordance with an embodiment of the present invention.
  • Figure 5 shows a STL network in accordance with an embodiment of the present invention.
  • Figure 6 shows a flow process in accordance with an embodiment of the present invention.
  • Figure 7 shows an illustration of temporal perception.
  • Figures 8(a)-8(d) illustrate four consecutive updates of a time adjacency matrix in accordance with an embodiment of the present invention.
  • Figure 9(a) illustrates a decay function in accordance with an embodiment of the present invention.
  • Figure 9(b) illustrates an update of the time-adjacency matrix using a decay-function in accordance with an embodiment of the present invention.
  • Figure 10(a) illustrates a multiple assignment of a sensed input in accordance with an embodiment of the present invention.
  • Figure 10(b) illustrates an update of the time- adjacency matrix using multiple assignment of a sensed input in accordance with an embodiment of the present invention.
  • Figure 1 l(a) illustrates a time-adjacency matrix.
  • Figure 1 l(b) illustrates a clustering of a time-adjacency matrix of Figure 11 (a) in accordance with an embodiment of the present invention.
  • Figure 12(a) illustrated a graph representing a time-adjacency matrix.
  • FIG. 12(b) illustrates the selection of nodes in the graph according in accordance with an embodiment of the present invention.
  • Figure 12(c) shows a grouping of a graph into three sub-graphs in accordance with an embodiment of the present invention.
  • Figure 12(d) shows a time adjacency matrix.
  • Figure 13 shows a flow process in accordance with an embodiment of the present invention.
  • Figure 14 shows at least a portion of a HTM-based system in accordance with an embodiment of the present invention.
  • Figure 15 shows a flow process in accordance with an embodiment of the present invention.
  • Figure 16 shows at least a portion of a HTM-based system in accordance with an embodiment of the present invention.
  • Figure 17 shows at least a portion of a HTM-based system in accordance with an embodiment of the present invention.
  • Figure 18 shows at least a portion of a HTM-based system in accordance with an embodiment of the present invention.
  • Figure 19 shows at least a portion of a HTM-based system in accordance with an embodiment of the present invention.
  • Certain aspects of the present invention include process steps and instructions described herein in the form of an algorithm. It should be noted that the process steps and instructions of the present invention could be embodied in software, firmware or hardware, and when embodied in software, could be downloaded to reside on and be operated from different platforms used by a variety of operating systems.
  • the present invention also relates to an apparatus for performing the operations herein.
  • This apparatus may be specially constructed for the required purposes, or it may comprise a general-purpose computer selectively activated or reconfigured by a computer program stored in the computer.
  • a computer program may be stored in a computer readable storage medium, such as, but is not limited to, any type of disk including floppy disks, optical disks, CD-ROMs, magnetic-optical disks, read-only memories (ROMs), random access memories (RAMs), EPROMs, EEPROMs, magnetic or optical cards, application specific integrated circuits (ASICs), or any type of media suitable for storing electronic instructions, and each coupled to a computer system bus.
  • the computers referred to in the specification may include a single processor or may be architectures employing multiple processor designs for increased computing capability.
  • the algorithms and displays presented herein are not inherently related to any particular computer or other apparatus.
  • Various general-purpose systems may also be used with programs in accordance with the teachings herein, or it may prove convenient to construct more specialized apparatus to perform the required method steps. The required structure for a variety of these systems will appear from the description below.
  • the present invention is not described with reference to any particular programming language. It will be appreciated that a variety of programming languages may be used to implement the teachings of the present invention as described herein, and any references below to specific languages are provided for disclosure of enablement and best mode of the present invention.
  • An "object” is at least partially defined as having some persistent structure over space and/or time.
  • an object may be a car, a person, a building, a word, a song, an abstract entity such as a concept or information flowing in a network.
  • an object in the world 10 may also be referred to as a "cause" in that the object causes particular data to be sensed, via senses 12, by a human 14.
  • the smell (sensed input data) of a rose (object/cause) results in the recognition/perception of the rose.
  • the image (sensed input data) of a dog (object/cause) falling upon a human eye results in the recognition/perception of the dog.
  • HTM hierarchical temporal memory
  • An HTM is a hierarchical network of interconnected nodes that individually and collectively (i) learn, over space and time, one or more causes of sensed input data and (ii) determine, dependent on learned causes, likely causes of novel sensed input data. HTMs, in accordance with one or more embodiments of the present invention, are further described below with reference to Figures 2 - 20.
  • An HTM has one or more levels of nodes. For example, as shown in Figure 2,
  • HTM 20 has three levels Ll, L2, L3, with level Ll being the lowest level, level L3 being the highest level, and level L2 being between levels Ll and L3.
  • Level Ll has nodes 22, 24, 26, 28; level L2 has nodes 30, 32, and level L3 has node 34.
  • the nodes 22, 24, 26, 28, 30, 32, 34 are hierarchically connected in a tree-like structure such that each node may have several children nodes (i.e., nodes connected at a lower level) and one parent node (i.e., node connected at a higher level). Note that it is also possible to have a single child node connected to multiple parent nodes.
  • Each node 22, 24, 26, 28, 30, 32, 34 may have or be associated with a capacity to store and process information.
  • each node 22, 24, 26, 28, 30, 32, 34 may store sensed input data (e.g., groups of patterns) associated with or derived from particular causes. Further, each node 22, 24, 26, 28, 30, 32, 34 may be arranged to (i) propagate information "forward” (i.e., "up” an HTM hierarchy) to any connected parent node and/or (ii) propagate information "backward” (i.e., "down an HTM hierarchy) to any connected children nodes.
  • sensed input data e.g., groups of patterns
  • each node 22, 24, 26, 28, 30, 32, 34 may be arranged to (i) propagate information "forward” (i.e., "up” an HTM hierarchy) to any connected parent node and/or (ii) propagate information "backward” (i.e., "down an HTM hierarchy) to any connected children nodes.
  • the nodes are associated or coupled to each other by links.
  • a link represents a logical or physical relationship between an output of a node and an input of another node. Outputs from a node in the form of variables are communicated between the nodes via the links.
  • Inputs to the HTM 20 from, for example, a sensory system are supplied to the level Ll nodes 22, 24, 26, 28.
  • a sensory system through which sensed input data is supplied to level Ll nodes 22, 24, 26, 28 may relate to commonly thought-of human senses (e.g., touch, sight, sound) or other human or non-human senses.
  • the range of sensed input data that each of the level Ll nodes 22, 24, 26, 28 is arranged to receive is a subset of an entire input space. For example, if an 8 x 8 image represents an entire input space, each level Ll node 22, 24, 26, 28 may receive sensed input data from a particular 4 x 4 section of the 8 x 8 image.
  • Each level L2 node 30, 32 by being a parent of more than one level Ll node 22, 24, 26, 28, covers more of the entire input space than does each individual level Ll node 22, 24, 26, 28. It follows that in Figure 2, the level L3 node 34 covers the entire input space by receiving, in some form, the sensed input data received by all of the level Ll nodes 22, 24, 26, 28. Moreover, in one or more embodiments of the present invention, the ranges of sensed input data received by two or more nodes 22, 24, 26, 28, 30, 32, 34 may overlap.
  • HTM 20 in Figure 2 is shown and described as having three levels, an
  • HTM in accordance with one or more embodiments of the present invention may have any number of levels.
  • the hierarchical structure of an HTM may be different than that shown in Figure 2.
  • an HTM may be structured such that one or more parent nodes have three children nodes as opposed to two children nodes like that shown in Figure 2.
  • an HTM may be structured such that a parent node in one level of the HTM has a different number of children nodes than a parent node in the same or another level of the HTM.
  • an HTM may be structured such that a parent node receives input from children nodes in multiple levels of the HTM.
  • HTM-based any entity that uses or is otherwise dependent on an HTM as, for example, described above with reference to Figure 2 and below with reference to Figures 3 - 19, may be referred to as an "HTM-based" system.
  • HTM-based system may be a machine that uses an HTM, either implemented in hardware or software, in performing or assisting in the performance of a task.
  • an HTM discovers one or more causes in its world from sensory input data received by the HTM.
  • an HTM does not necessarily have a sense particular to each of the types of causes being sensed; instead, an HTM may discover from raw sensed input data that causes such as cars and words exist. In such a manner, an HTM is able to learn and form representations of causes existing in its world.
  • an "object" has persistent structure. The persistent structure causes persistent patterns to be sensed by an HTM. Each sensed input pattern has one or more spatial attributes. In other words, each sensed input pattern may be thought of as being represented as a particular set of bits.
  • a node in an HTM "learns," i.e., stores and associates with a common cause, sensed input patterns by determining cooccurrences of sensed input patterns in its input. Determining co-occurrences of sensed input patterns involves determining which sensed input patterns are active at around the same time at a rate statistically greater than what would be expected based on mere chance. For example, if an HTM node having one hundred inputs has seven inputs that become active together at some statistically significant rate, then the HTM node learns that the sensed input patterns at those seven inputs co-occur.
  • an HTM node may not be necessary for an HTM node to learn all sensed input patterns occurring together at some statistically significant rate.
  • an HTM node may store the x most common sensed input patterns found in its input.
  • an HTM node may store sensed input patterns according a pre-defined measure of significance (e.g. uniqueness).
  • learning causes in an HTM-based system involves learning patterns and groups of patterns.
  • patterns and groups of patterns that co- occur frequently are stored and assigned to the same causes.
  • groups of patterns that co-occur frequently at some statistically significant rate may be assigned to the same cause.
  • learning causes may effectively entail mapping many patterns and/or inputs to a single cause.
  • Such assigning of multiple patterns and/or inputs to a single cause may also be referred to as "pooling.”
  • Spatial group learning attempts to learn a set of meaningful co-occurrences in a set of input patterns. Spatial pooling is dependent on "spatial" similarities between two or more patterns (noting that a pattern may actually represent a sequence from a lower level). In embodiments which use spatial pooling, an HTM node may compare a spatial property of a received sensed input pattern with that of a learned sensed input pattern (or "quantization" point). If the two patterns are "similar enough” (i.e., have enough “overlap"), then the received sensed input pattern may be assigned to the same cause as that of the quantization point.
  • a quantization point is equal to ' 10010110'
  • a received sensed input pattern of ' 10011110' may be assigned to the same cause as that of the quantization point due to there being a difference of only one bit between the two patterns.
  • the amount of similarity needed to perform such "spatial" pooling may vary within and/or among HTM-based systems.
  • Temporal groups are learned using temporal pooling algorithms that recognize sensed input patterns that occur together at a rate statistically greater than what would be expected based on mere chance. For example, if of fifty sensed input patterns learned by an HTM node, three occur in order at some statistically significant rate, then the HTM node may learn that group of sensed input patterns
  • One type of temporal pooling involves assigning patterns to the same cause based on temporal adjacency. For example, if an HTM node receives pattern A followed by pattern B followed by pattern D, then patterns A, B, and D may be assigned to the same cause as there is some likelihood that this sequence of patterns was caused by the same object. Accordingly, such "temporal" pooling enables the mapping of patterns, some or all of which may have no significant spatial overlap, to a single cause.
  • the temporal groups learned by a HTM node may each be represented by one or more variables. As each learned temporal group is associated with a particular cause, each variable accordingly represents a different cause.
  • the HTM node may pass each of its variables up to a parent node via a vector containing probabilities as to the likelihood that each of its learned groups is active at its input at a given time.
  • the parent node may then (i) determine temporal adjacencies of its sensed input patterns (i.e., the variables received from its child node), (ii) learn sensed input patterns as described above, and (iii) learn high-order groups of learned sensed input patterns (i.e., learn groups of temporally co-occurring variables representing groups learned by its child node).
  • a spatio-temporal learning (STL) node is comprised of one or more HTM nodes which learn both spatial and temporal groups of sensed input patterns.
  • STL nodes learn spatial and temporal groups of inputs over time.
  • Figure 3 illustrates a STL node 300 according to one embodiment.
  • the STL node 300 comprises a spatial pooler 302 which is used to determine spatial groups in the set of sensed input patterns.
  • the STL nodes 300 further comprises a temporal pooler 304 used to determine groups of sensed input patterns that temporally co-occur.
  • each STL node 300 may contain additional programs which govern the operation of the spatial pooler and temporal pooler based on aspects of the node.
  • the spatial pooler 302 receives some input patterns 350. Generally, the spatial pooler 302 identifies spatial co-occurrences among input patterns. At each time-step, the spatial pooler 302 outputs a set of probabilities P(e " t
  • S) represents the probability of observing e " (evidence or input patterns from a lower level) at time t when over learned spatial co-occurrences S.
  • S) represents the probability of observing e " (evidence or input patterns from a lower level) at time t when over learned spatial co-occurrences S.
  • S) represents the probability of observing e " (evidence or input patterns from a lower level) at time t when over learned spatial co-occurrences S.
  • S) 370 are communicated outside of the spatial pooler 302. Further, note that the spatial pooler 302 , has a behavior (i.e., outputting distributions over learned spatial groups of sensed input patterns) that may be dependent on the type and/or topology of network of which the spatial pooler 302 is part.
  • the temporal pooler 304 Based on the set of probabilities P(e " t
  • G) 390 corresponds to a different learned temporal group G 1 .
  • the learned temporal groups themselves may not be communicated outside of the temporal pooler 304.
  • the temporal pooler 304 has a behavior (i.e., outputting distributions over learned groups of sensed input patterns) that may be dependent on the type and/or topology of network of which the group learner 44 is part.
  • a STL node 300 may comprise more than one HTM node.
  • a STL node 300 may comprise one HTM node executing a spatial pooler 302 which outputs a set of probabilities P(e ⁇ t
  • a STL node 300 may comprise one or more HTM nodes executing spatial poolers 302 which output one or more sets of probabilities P(e ⁇ t
  • a STL node 300 may comprise multiple HTM nodes executing spatial poolers 302 and a single HTM node executing a single temporal pooler 304.
  • FIG. 4 provides an illustration of a STL top node 400.
  • a STL top node 400 is a node in the HTM network which is at the top of the hierarchy of nodes and directly or indirectly receives inputs from the all of the nodes in the HTM network.
  • the STL top node 400 receives a set of inputs 410 that represent a set of probabilities P(e " t
  • a STL top node 400 comprises a spatial pooler 302 which produces a set of probabilities P(e ⁇ t
  • the STL top node 400 further comprises a Supervised Mapper 460 which receives the set of P(e " t
  • C) 480 is used to assign a category C to a sensed input pattern based on the sensed input pattern with the highest probability P(e ⁇ t
  • the Supervised Mapper 460 may alternatively output category C assigned to the sensed input pattern.
  • a Spatio-Temporal Learning (STL) network is a hierarchical network including a plurality of STL nodes 300 and a STL top node 400.
  • a STL network may use different spatial and temporal group algorithms at different levels of the hierarchy.
  • Fig. 5 illustrates a STL network 500 in which different temporal and spatial pooling algorithms are used at different levels.
  • the STL nodes 512, 514, 516, 518 receive inputs from a data sensor.
  • the data sensor partitions input data and distributes the input data to the first level Ll nodes.
  • the data sensor may partition a picture into 4 equal parts and distribute them to the first level Ll nodes 512, 514, 516 and 518.
  • the STL nodes in the STL network 500 execute different Spatial Pooler 302 algorithms at each level.
  • the STL nodes execute Gaussian Spatial Poolers 515 that are described in detail below.
  • the nodes 522, 524, 532 execute Dot or Product Spatial Poolers 525 that are described in detail below.
  • STL nodes at different levels of a STL network 500 may execute different or additional spatial pooling algorithms such as Sparse Distributed Code Poolers that are described in detail below.
  • nodes at different levels of a STL network may execute different or additional temporal pooling algorithms.
  • the STL top node further executes a Supervised Mapper 460 used to generate a probability that a sensed input belongs to a category of inputs.
  • the Spatial Poolers 302 function in both learning and inference. During learning, the Spatial Pooler 302 functions to identify common cooccurrences between patterns of inputs. According to the embodiment, these patterns of inputs may be represented in different ways, e.g., as input vectors.
  • the Spatial Pooler 302 generates the set of probabilities P(e ⁇ t
  • the Spatial Pooler 302 can be implemented in a variety of ways.
  • Four different spatial pooling algorithms for STL nodes 300, 400 are disclosed herein. These spatial pooling algorithms are referred to as the Gaussian, Dot, Product and Sparse Distributed Code Pooling algorithms.
  • the optimal spatial pooling algorithm depends primarily on the nature of the input vectors. In one embodiment, Gaussian spatial pooling is optimally used for nodes that receive input directly from a sensor. In one embodiment, Dot or Product spatial pooling is optimally used for middle or top nodes that receive input from other STL nodes 300, 400.
  • the Gaussian spatial pooling algorithm performs a form of vector quantization, which is useful for sensory input signals, whereas Dot/Product pooling does not.
  • Sparse Distributed Code Poolers are used to learn canonical inputs or spatial co-occurrences from input patterns which may contain more than one independent canonical input. Sparse Distributed Code Poolers decompose the sensed input pattern received from a lower node into "sparse distributed codes" or spatial co-occurrences of independent canonical inputs.
  • the Gaussian spatial pooling algorithm performs a form of vector quantization, which is useful for sensory input signals, whereas Dot/Product pooling does not.
  • the Dot and Product algorithms process the distribution sent by each child STL Nodes by finding the index of the element which represents the "winning" temporal group.
  • the Spatial Pooler 302 analyzes the stream of sensed input patterns in order to generate a spatial co-occurrence matrix.
  • This spatial co-occurrence matrix represents the set of all possible input patterns as a relatively small, finite set of canonical input patterns.
  • These canonical input patterns selected by the spatial pooler are referred to herein as "spatial co-occurrences" because they represent distinctive occurrences of input patterns from the node's children that tend to appear together far more often than would be expected by random chance.
  • the spatial co-occurrence matrix W is a list of spatial co-occurrences. Each spatial co-occurrence is stored as a row Wj, and W contains one column for each element in the input pattern.
  • each node's input pattern is a row vector that is formed by concatenating together the output vectors of each of the node's children.
  • the spatial co-occurrence matrix starts out empty. When the Spatial Pooler 302 selects a particular input vector to be a spatial co-occurrence, it simply appends this input vector to W as a new row.
  • the Spatial Pooler 302 also counts the number of input vectors that are pooled with each of the spatial co-occurrences in the matrix in addition to its primary task of building a spatial co-occurrence matrix. These frequency counts are stored in a single vector called counts. Upon the completion of learning, this vector will contain one integer for each spatial co-occurrence in W. The value of theyth element is simply the total number of input vectors that were pooled with they ' th spatial cooccurrence throughout the learning stage. The Spatial Pooler 302 does not use this frequency count information directly, but instead provides it to the Temporal Pooler 304. [0074] In the inference stage, the Spatial Pooler 302 no longer updates the spatial cooccurrence matrix W, and instead compares each new input vector to the spatial cooccurrences in W.
  • the Spatial Pooler 302 computes a "belief vector" y for its input vector x.
  • This output vector is a distribution over spatial co-occurrences, so it contains one element for each row in W.
  • the output vector y is handed off to the Temporal Pooler 304.
  • the Spatial Pooler 302 can be thought of as a pre-processor for the Temporal Pooler 304.
  • the Spatial Pooler 302 simplifies the inputs to the Temporal Pooler 304 by pooling the vast space of input vectors into a relatively small set of discrete spatial cooccurrences that are easier to handle.
  • belief represents a generalized measure of the likelihood that a particular input vector x and a particular spatial co-occurrences w both represent the same underlying real-world cause. Although this use of "belief does not carry a precise mathematical meaning, a higher value in the belief vector implies a higher likelihood that the input vector and the corresponding spatial co-occurrences were produced by the same cause.
  • the Gaussian Spatial Pooler 515 receives a stream of input vectors. For each received input vector, the Gaussian Spatial Pooler 515 indentif ⁇ es the particular input vector x as a new spatial co-occurrence if x is sufficiently dissimilar (by a Euclidean distance metric) to all of the spatial co-occurrences that the Gaussian Spatial Pooler 515 has previously identified. The Gaussian Spatial Pooler 515 uses a Euclidean distance metric to compare the input vectors presented during training. The squared Euclidean distance between an input vector x and an existing spatial co-occurrence w is computed as follows:
  • the Gaussian Spatial Pooler 515 identifies new spatial co-occurrences based on a specified threshold distance value, maxDistance. If the squared distance between an input vector x and an existing spatial co-occurrence w is less than maxDistance, the input vector is not considered to be a new spatial co-occurrence and is pooled together with that existing spatial co-occurrence . In this case, the frequency count of the existing spatial cooccurrence w in the counts vector is incremented by 1. If the input vector's squared distance to all existing spatial co-occurrences is greater than maxDistance is the input vector identified as a new spatial co-occurrence, in which case the Gaussian Spatial Pooler 515 appends the new spatial co-occurrence to W.
  • maxDistance a specified threshold distance value
  • the Gaussian Spatial Pooler 515 algorithm is highly dependent on the value of maxDistance.
  • the Gaussian Spatial Pooler 51 5 will identify more spatial co-occurrences when maxDistance is low. A high value of maxDistance will result in fewer spatial cooccurrences. If maxDistance is equal to 0, every unique input vector presented during training will be selected as a spatial co-occurrence.
  • the Dot or Product Spatial Poolers 525 operate in an identical manner during the learning stage and only differ in operation during the inference stage.
  • the outputs from the node's children are concatenated together into a single row vector representing an input pattern.
  • the Dot/Product Spatial Pooler 525 tracks which elements of this concatenated vector came from which child in the STL network 500, using this information to make spatial pooling decisions.
  • the Dot/Product Spatial Poolers 525 employ a winner-take-all approach to spatial pooling.
  • the STL nodes 300, 400 When presented with a concatenated input vector x, the STL nodes 300, 400 divide this input vector into NcMdren portions, where NcMdren denotes the number of children feeding into the STL node 300, 400. Each of these portions contains the elements of x that were contributed by a particular child. [0079] The elements in each portion represent a distribution over the temporal groups of a single child. Within each portion, the Dot/Product Spatial Pooler 525 locates the largest element and declares it the "winner". The Dot/Product Spatial Pooler 525 then updates this portion of x by setting the winning element's value to 1 and setting all the other elements in that portion to 0.
  • the Dot/Product Spatial Pooler 525 performs this operation on each portion of the input vector.
  • the result is an updated input vector containing NcMdren elements with a value of 1, with all remaining elements set to 0.
  • Each child's portion contains a single 1.
  • the Dot/Product Spatial Pooler 525 inserts the updated input vector into the spatial co-occurrence matrix W if the vector is not already present in the matrix. After the node sees the entire set of training data, W will be a list of all unique input vectors as updated by the winner-take-all rule.
  • the Dot/Product Spatial Pooler 525 maintains a count of the number of input vectors that have been pooled with each spatial co-occurrence. If a particular input vector is found to match an existing spatial co-occurrence, the frequency count associated with that spatial co-occurrence is incremented.
  • the Dot/Product Spatial Pooler 525 requires no parameters. Despite the fact that the Dot/Product Spatial Pooler 525 stores as a spatial co-occurrence every unique input vector (subsequent to winner-take-all processing), the memory usage is not typically prohibitive in practice. Real-world data tends to have a great deal of causal structure; therefore, the number of unique spatial co-occurrences encountered each STL node 300, 400 is far lower than what would be expected if random inputs were presented to the network. In some embodiments the number of unique spatial co-occurrences encountered by a Zeta -1 node 300, 400 can still become large. In these embodiments, the Dot/Product Spatial Pooler 525 uses sparse matrices and other coding optimizations to reduce memory usage during the learning stage.
  • the Spatial Pooler 304 receives an input vector x and computes a new distribution vector y with one element for each spatial co-occurrence in the spatial co-occurrence matrix W.
  • Each of these values yj represents the belief that the input x belongs to the spatial co-occurrence "bin" W].
  • a belief value generated by the Spatial Pooler 304 represents neither a true probability nor a mathematically rigorous likelihood. Instead, a belief value yj is best thought of as a generalized measure of "belief that the input vector x was generated by the underlying cause associated with the spatial cooccurrence Wj from the spatial co-occurrence matrix.
  • the Gaussian Spatial Pooler 515 receives an input vector x, it first computes the Euclidean distance between x and each of the spatial co-occurrences in W. The Gaussian Spatial Pooler 515 then converts these distances to belief values using a Gaussian approach.
  • the algorithm models the input vector x as a random sample drawn from one o ⁇ Ncoincs Gaussian probability distributions. Each of these distributions is centered on a different spatial co-occurrence vector in W; the Ncoincs rows of W represent the mean vectors of the Gaussian distributions.
  • the distributions all have an identical variance, which is uniform across all dimensions of the input space.
  • the Gaussian Spatial Pooler 515 computes the belief that x was generated from the cause represented by the f h spatial co-occurrence as the local probability density of the / h Gaussian distribution, as measured at point x:
  • the belief j; calculated by the Gaussian Spatial Pooler 515 would be strictly proportional to the probability that the input vector x was drawn from the 7 th distribution.
  • this model provides an approximation, and produces an exponential decay of the belief in the 7 th cause with increasing distance between x and Wj.
  • the aforementioned parameter sigma controls the rate at which the Gaussian probability distributions decay with distance. A low value of sigma will result in very small Gaussian "clouds" centered tightly around each spatial co-occurrence. Only the spatial co-occurrences that are highly similar to the input vector will receive non-infinitesimal belief values.
  • Dot and Product algorithms store input vectors updated using the winner-take- all approach. Each of these vectors has NcMdren elements set to 1 — one such element for each child node — and all other elements set to 0.
  • the Dot inference algorithm receives a concatenated input vector from its child nodes, it generates a belief value for each of its stored spatial co-occurrences using a dot product operation.
  • the belief value y 3 associated with they ' * spatial co-occurrence is simply the dot product of the / h (updated) spatial cooccurrence vector with the input vector:
  • a spatial co-occurrence vector defines a single "winning" element for each of its children
  • the input values for each of these winning elements are simply summed together to yield the belief value.
  • the Dot Spatial Pooler 525 repeats this process for each of the spatial co-occurrences in W to produce the output vector y with N CO ⁇ ncs elements. Because the dot product operation reduces to a summation, the Dot Spatial Pooler 525 provides flexibility in instances in which most of the inputs from a node's children match a particular spatial co-occurrence well, but one or two children are badly mismatched. In this case, the high input values of the matching child nodes will dominate the summation and tend to compensate for the low input values coming from the mismatched child nodes.
  • the Product Spatial Pooler 525 inference algorithm is very similar to the Dot algorithm.
  • the Dot Spatial Pooler 525 computes a belief value, it sums the support from each of its children.
  • the Product Spatial Pooler 525 multiplies the support. This simple distinction generates to very different belief vectors.
  • a node with 10 children If the node uses the Dot Spatial Pooler 525, a spatial co-occurrence that matches well with the input vectors from nine of the ten children but is dissimilar to the input from the tenth child will still receive a very high belief value. But if the node uses the Product Spatial Pooler 525, the tenth child can "veto" the spatial co-occurrence by zeroing out the overall belief value. Accordingly, the Product Spatial Pooler 525 algorithm thus requires strong consensus among the children to produce a non-infinitesimal belief value.
  • a STL network may be composed entirely of STL nodes 300, 400 which
  • Sparse Distributed Code Poolers as their Spatial Pooler 302. In other embodiments, only lower level or higher level nodes use Sparse Distributed Code Poolers as their Spatial Pooler
  • Sparse Distributed Code Poolers decompose a set of received input patterns into a set of spatial co-occurrences representing "sparse distributed codes".
  • Sparse distributed codes are components of an input pattern which are independent and thought to have the same cause.
  • the components in the set of sparse distributed codes are "sparse", meaning that Sparse Distributed Code Pooler learns only the minimal number of independent components that can be used to represent the sensed input patterns.
  • Sparse Distributed Code Pooler learns only the minimal number of independent components that can be used to represent the sensed input patterns.
  • different variations of known sparse distributed code learning algorithms can be used to identify the optimal independent components within the sensed input patterns. These algorithms can include Fouldeac learning algorithms or any other type of Hebbian learning algorithms. Other sparse distributed code learning algorithms will be well known to those skilled in the art.
  • Sparse Distributed Code Poolers are well suited for use as Spatial Poolers 302 in HTM networks as they are unsupervised learning algorithms. In these algorithms, the Sparse Distributed Code Pooler learns the set of spatial co-occurrences or "sparse distributed codes" by iteratively reweighting an initial set of components. Once these components have been identified, the Sparse Distributed Code Pooler can either switch into inference mode in which the Sparse Distributed Code Pooler determines a set of probabilities P(e " t
  • the Sparse Distributed Code Pooler can continue to both determine the set of probabilities P(e " t
  • the Temporal Pooler 304 learns groups and builds a time-adjacency matrix which is updated over time ST50. Further, the group learner is arranged to collect statistics from the time-adjacency matrix to learn temporal groups ST52 and then, based on its learned groups and statistics thereof, compute probability distributions for new inputs (as described above) ST54.
  • Figure 7 illustrates an example of sensing or perceiving a right corner moving in and out of vision.
  • the individual sequences of input patterns may represent the same input of a right corner, the sequences of viewing the corner moving into view may be interspersed over time. Additionally, not all of the sequences of patterns representing the right corner are comprised of the exact same image or pattern but instead patterns which represent the same cause.
  • one limitation of temporal pooling based strictly on sequential patterns is that this technique often causes a large number of sequences to be created and grouped. When inputs or patterns of inputs with the same cause are sensed in real life, it is not always the case that they adhere to the same sequence every time.
  • Random noise and variety in the order which inputs or patterns are sensed can create a large number of sequences of events, which all have the same cause. Due to different orders in which inputs can be sensed, a large number of patterns which have the same cause can create a combinatorial explosion when considering all possible sequences the patterns can form. Enumerating all of the possible different sequences is not only computationally expensive but often inaccurately reflects the frequency with which patterns co-occur. Promiscuous sequences can sometimes lead to an even or "flat" distribution in frequencies of temporal adjacencies. This "flat" distribution can occlude the detection of differences in frequency of temporal adjacency due to a common cause.
  • Group based temporal pooling addresses these problems in temporal and spatial pooling by identifying groups of input patterns thought to have the same cause by identifying temporally temporal adjacency between inputs without relying on the sequence of occurrence of these inputs.
  • Temporal adjacency is used herein to refer to two input patterns occurring at the exact same time as well as close in time point to each other such as directly before or after each other (temporally adjacent) or close to each other in a sequence of inputs over time (temporally proximal).
  • the order in which the input patterns occur can be symmetrical or asymmetrical.
  • the identified temporal adjacency values may be based or the order the inputs occur (e.g. A..
  • the groups identified by group based temporal pooling are unordered, meaning that there is no inherent structure or sequence in which the input patterns occur within the group.
  • the input patterns may be partially ordered in a group based on various criteria such as frequency of occurrence.
  • temporal adjacency is enumerated for each of the input patterns in a time-adjacency matrix.
  • Each row in the time adjacency matrix represents a sensed input pattern for a time point.
  • Each column in the time adjacency matrix also represents a sensed input pattern.
  • the matrix is enumerated to reflect the frequency of temporal adjacency.
  • the matrix is updated at the row corresponding to the input at time t and the column corresponding to the input that occurred at the previous time point W t -i.
  • the matrix is also updated at several previous time points.
  • the number of time points can range from 1 to n, where n represents all previous all time points used to update the matrix.
  • n may be parameterized or take a default value.
  • the matrix may be updated with an integer number which enumerates the number of times the input at time t and the input at time W t -i are temporally adjacent, or any alternative numeric variable to represent the frequency of temporal adjacency. Additionally, alternate data structures may be used to score temporal adjacency. Various combinations of data structures and numeric variables to enumerate frequency will be apparent to those skilled in the art.
  • a decay function can be used to weight the enumeration of temporal adjacency.
  • the decay functions may be linear in proportion to n, which represents the number of time points between t and W t . n .
  • the decay function may be exponential or pseudo- exponential over time. Examples of suitable decay functions to weight temporal adjacency by proximity include but are not limited to: persistence decay functions, block decay functions and exponential decay functions.
  • the decay function may be adjusted to weight temporally proximal patterns with a greater or lesser weight.
  • Figures 9(a) and 9(b) illustrate a linear decay function using the integer based enumeration shown in Figures 8(a)-(d).
  • Figure 9(a) illustrates a decay function which is linearly proportional to n, where n equals 4.
  • the maximum number used to enumerate the temporal adjacency is 4 for W t -i.
  • Each pattern at the previous time points (W t -2, W t - 3 , W t - 4 ) is assigned an integer value which reflects its temporal distance relative to the pattern at time t.
  • the linear decay function subtracts one integer value for every time point between the enumerated pattern and the pattern at time t.
  • Figure 9(b) illustrates one update of the time adjacency matrix using the decay function.
  • the row representing the pattern at time t is updated at the columns representing patterns at Wt-i, Wt-2, Wt-3 and W t . 4 with values 4, 3, 2 and 1, respectively.
  • the enumeration of the time adjacency matrix is also weighted to reflect spatial proximity of the patterns in some embodiments. In systems where a large number of potential patterns may exist, it may be necessary to represent new patterns using a set of quantization points or by defining a threshold on the number of input patterns to include in the time-adjacency matrix. As discussed above, quantization points may be determined based upon frequency of occurrence.
  • Quantization points may also be determined using standard machine learning techniques such as principle components analysis. Alternatively, quantization points may be determined using other unsupervised learning methods such as Gaussian clustering. According to the embodiment, the number of quantization points may be parameterized by specifying the number of input patterns to include in the time-adjacency matrix.
  • each pattern may be assigned a set of features and similarity may be determined between two patterns using a variety of distance functions.
  • the set of features and distance functions may be specific to the type of pattern used and the application of the pooled groups. For example, patterns comprised of images may be compared using distance functions based on image-based features such as pixel color, luminescence, etc. Patterns comprised of sequences may be compared using distance functions including mismatch values and gapped alignments. Additional similarity metrics and distance functions will be readily apparent to those skilled in the art in light of this disclosure.
  • the new pattern may be assigned to multiple different existing inputs to represent spatial proximity. Multiple assignment of new inputs is used to enumerate several existing patterns based one new input, instead of creating a new row in the time-adjacency matrix representing every input not previously sensed by the system.
  • the temporal adjacency enumeration value may be multiply divided over several images in any suitable method. If similarity is probabilistically determined using machine learning methods, multiple assignment may correspond to the probability the two images are the same. Alternately, multiple assignment may be based on one or more similarity values.
  • FIG. 10(a) The concept of multiple assignment is illustrated in Figure 10(a).
  • the image of the intersection of perpendicular horizontal and vertical lines has both high similarity to the images of the horizontal line and the image of the vertical line.
  • Both the images of the horizontal line and the vertical line are assigned a value of 1 A or .5 of a temporal adjacency enumeration value of 1.
  • the rows corresponding to both the horizontal line image and the vertical line image are updated at time t at the column corresponding to the image of time W t -i-
  • Figure 1 l(a) illustrates a simple example of a complete time- adjacency matrix with two distinct groups of input patterns, a set of right corners and a set of left corners.
  • the only pattern (i.e. image) that has overlap between the two groups is that of the straight line, which temporally co-occurs or is adjacent to both the left corner and the right corners. If the pattern representing the straight line is removed from the time adjacency matrix, the right corners and left corners would form two separate and distinct groups.
  • these groups are identified using hierarchical agglomerative clustering or any method of "hard” clustering wherein each row of the time-adjacency matrix is assigned to only one cluster or group.
  • the clustering algorithm may learn a number of clusters reflective of the natural groups of input patterns. Alternatively, the clustering algorithm may take a number of clusters to form as a parameter.
  • the identification of the two groups of input patterns through clustering is illustrated in Figure 1 l(b). Suitable clustering methods include single linkage clustering and n-means clustering. In this type of clustering, the temporal adjacencies enumerated in each row of the matrix form the set of values used to pair-wise similarities between the two input patterns. Any type of distance metric can be employed to determine the pair-wise similarities which are then used to cluster or group the rows representing the input patterns.
  • the groups are identified though a method which employs a greedy graph traversal to identify cliques or groups of co-occurring patterns. This method relies on the principle of graph connectivity to identify and remove patterns which have overlap between multiple groups. This concept is illustrated in Figures 12(a)-(c).
  • Figure 12(a) depicts a graphic illustration of a time adjacency matrix wherein each pattern is represented by a node in the graph and each edge between the nodes in the graph indicates temporal adjacencies between the two patterns.
  • the two center nodes are identified as a distinct group of having high connectivity to other nodes.
  • Each node is connected to four other nodes, causing both of the center nodes have a high sum of temporal adjacency values due to increased connectivity. However, the two center nodes have low temporal adjacency values with the other nodes, having the largest temporal adjacency value with each other.
  • the identified group of center nodes (Group 0) is removed from the graph, forming two other distinct sub-graphs representing the remaining groups (Group 1 and 2).
  • This method of identifying groups based on the time-adjacency matrix works well because there are several patterns which will be over-represented or sensed more frequently than others, leading to both higher connectivity and also higher likelihood of connectivity (i.e. a high number of temporal adjacencies) with input patterns in multiple groups. Since the rows representing these frequently sensed patterns will be enumerated at each time point, the row representing pattern will have a large value when summed. The inequity in frequency of different patterns produces an uneven distribution over the sum of the rows of the matrix. The sum of the rows of the matrix can then serve as an indicator of the relative frequency of temporal adjacency or "connectivity" of each pattern.
  • a greedy algorithm is used to identify groups based on the above discussed connectivity as an indicator of group overlap.
  • the greedy algorithm works by the principle of identifying for each sensed input pattern the nearest neighbors, that is, the sensed input patterns which have the highest frequency of temporal adjacency based on the time-adjacency matrix.
  • the group is defined as the smallest set of elements in which every element's closest topN neighbors are in the group.
  • the topN neighbors are the highest N neighbors according to the sorted rankings representing temporal adjacency, where N is a numeric value.
  • the greedy algorithm may be parameterized to allow the user to specify the N number of nearest neighbors of each element to be contained in each group.
  • the topN parameter may have as default value of 1.
  • Other parameters to allow the user to control the greedy algorithm include maxGroupSize and minGroupSize.
  • maxGroupSize allows the user to specify a maximum number of elements in a group and provides the user the ability to ensure that not all elements will be grouped into one group.
  • minGroupSize allows the user to specify a minimum number of elements to be contained within each group, thereby allowing the user to control spurious small groups.
  • Figure 13 illustrates a flow control of the greedy algorithm according to one embodiment.
  • the time-adjacency matrix is sorted by rowSum values in order to rank the rows representing different inputs according to their frequency of occurrence ST60.
  • the element with the highest rowSum is selected and added to the group ST62.
  • the neighbors of the first selected element are identified in the time adjacency matrix by determining which elements correspond to enumerated columns with in the row of the selected element ST64. The identified neighbors are checked to determine if all of the neighbors have already been selected ST66.
  • the identified neighbors are then sorted according the numeric value in of their corresponding columns in the time-adjacency metric which represents the temporal adjacency of the each neighbor relative to the first selected element.
  • a number of elements equal to the lesser of the number of neighbors remaining unselected or the topN neighbors is then selected for addition to the group ST68. If the number of elements in the group after addition is less than maxGroupsize the method is repeated with all the elements of the group ST70.
  • a tie breaking heuristic is used to determine which element to add to the group.
  • the original ranking of rowSum is used to determine the element with the highest frequency of occurrence of the two elements and add it to the group. For example, if the value of topN is 3 and the respective values of the top 4 neighbors are ⁇ 9,8,4,4 ⁇ then the neighbor with the value 4 with the higher rowSum is added. [0110] If all of the neighbors of the elements are already selected, the unselected element with the highest rowSum is selected ST62, it's neighbors are identified ST64, and subsequent steps may be repeated until all of the unselected elements have been selected and grouped.
  • the Temporal Pooler 304 receives a belief vector y representing the set of probabilities P(e " t
  • the temporal pooler uses its groups and matrix of weights to calculate a distribution over groups P(e ⁇ t
  • a Supervised Mapper 460 receives Spatial Pooler 302 as input. During supervised learning, the Supervised Mapper 460 receives the index of a winning spatial co-occurrence from the Spatial Pooler 302 in conjunction with a category label. During inference, it receives the belief vector y representing the set of probabilities P(e " t
  • the Supervised Mapper 460 forms a matrix with
  • Ncoinc rows and one column for each category C it receives as input.
  • the Supervised Mapper 460 retrieves an index for the received category C on each iteration and increments the corresponding value in its mapping matrix.
  • the mapper receives the belief vector y from the Spatial Pooler 302, and it produces a distribution over categories P(e ⁇ t C), which are the output of the node.
  • an HTM may determine causes of novel input using what may be referred to as "inference.”
  • an HTM may infer which of its learned causes is/are the source of the novel sensed input data based on statistical comparisons of learned groups and patterns with patterns in the novel sensed input data.
  • a cause is determined based on a single sensed input.
  • An HTM node receives the newly sensed input, the HTM node assigns probabilities as to the likelihood that the new sensed input matches each of its learned sensed input groups. Then, as described above, the distribution over the set of groups learned by the HTM node is passed to a higher level node.
  • the distribution passed by an HTM node is derived from a "belief as to the likelihood that each learned cause is the cause of sensed input patterns at the input of the HTM node.
  • a "belief also includes those messages that are derived from or based on the belief. For example, an HTM node having learned five causes may deterministically assign scores that represent the probability that each of the five learned causes is the cause of sensed input.
  • the inputs are not exclusively associated with a single cause. Therefore, in some instances one or more scores indicating the probability that learned causes are the cause of the sensed input could equal 1 or a corresponding score which indicates a high likelihood that the learned cause is the cause of the sensed inputs. These scores (or "beliefs" as described above) may be normalized (or un-normalized) and passed to a parent node.
  • the parent node may then determine temporal adjacencies among the distributions sent from its child nodes, and then, based on its learned sensed input patterns and groups thereof, pass to a yet higher level node its own belief as to the likelihood that each of its learned causes is the cause of sensed groups at its input. In other words, a parent node forms its own "higher level” belief as to the cause of the sensed input patterns at least partly based on some statistical convergence of the beliefs passed from its child nodes.
  • one or more prior probabilities may be set manually in addition to or instead of having prior probabilities set via prediction. In other words, an HTM may be manually controlled to anticipate a particular cause or set of causes.
  • inferring causes of sensed input patterns involves passing beliefs from lower level nodes to higher level nodes.
  • such "belief propagation" is shown in HTM 80 (beliefs indicated with arrows; nodes shown, but not labeled).
  • a belief is a vector of values, where each value represents a different cause.
  • a current belief of a node may be a distribution of several causes being at least partially active at the same time.
  • the values in the belief vector may be normalized so that a stronger likelihood of one cause represented in the vector will diminish the likelihood of other causes represented in the vector.
  • a meaning of a value representing a cause in a belief vector may not vary depending on what other causes represented in the belief vector are active.
  • a Zeta -1 network is a hierarchy of connected nodes. Each node may be thought as having a belief. In one or more embodiments of the present invention, a belief at one node may influence a belief at another node dependent on, for example, whether the nodes are connected via a conditional probability table (CPT).
  • CPT conditional probability table
  • a CPT is a matrix of numbers, where each column of the matrix corresponds to the individual beliefs from one node, and where each row of the matrix corresponds to the individual beliefs from another node.
  • a higher level node may form a belief about precipitation and have values representing the likelihood of the following causes: “sunny”; “rain”; “sleet”; and “snow”.
  • the belief about air temperature in the lower level node may inform the belief about precipitation in the higher level node (and vice-versa).
  • multiplying the vector representing the belief about air temperature in the lower level node by the CPT results in a vector representing the appropriate belief about precipitation in the higher level node.
  • belief propagation allows an HTM to infer causes such that each node in the HTM represents a belief that is maximally or optimally consistent with its input. Note that performing inference in such a manner results in ambiguities being resolved as beliefs ascend through the HTM.
  • the parent node may decide with relatively high certainty that a "cat” is present and not a "dog” or "pig.” The parent node effectively settled on “cat” because this belief is the only one that is consistent with its inputs, despite the fact the "cat" image and the "cat” sound were not the most likely beliefs of its child nodes.
  • a higher level node in an HTM may pass a
  • the "prediction” is a "belief in that it contains values representing the likelihoods of different causes.
  • the vector representing the belief in the higher level node may be multiplied by an appropriate CPT to inform a belief in the lower level node.
  • a higher level node in an HTM uses its learned sequences combined with recent state information (i.e., the current input to the higher level node) to (i) predict what its next belief should be and (ii) then pass the expectation down to one or more lower level nodes in the HTM.
  • Figure 15 shows a flow process in accordance with an embodiment of the present invention. Particularly, Figure 15 shows in summary the steps of belief propagation described above.
  • a current node in the HTM receives input (in the form of sensed input patterns or beliefs from lower level nodes) ST82. Based on the received input and any beliefs passed down from a higher level node, the current node forms/adjusts its belief as to the likelihood of causes at its input distributed over its learned causes ST84. This belief is then passed to higher level and/or lower level nodes to inform beliefs at those nodes ST86.
  • an HTM network may be provided as a software platform.
  • an HTM network (nodes shown, but not labeled) 164 may run across several CPUs 166, 168, 170.
  • the CPUs 166, 168, 170 may either be part of a single system (e.g., a single server) or multiple systems.
  • an HTM network may be created in software across several multiprocessor servers, where such a group of servers may be referred to as a "cluster."
  • the servers in a cluster may be heterogeneous, i.e., the servers may have differing configurations/specifications (e.g., clock speeds, memory size, number of processors per server).
  • the servers may be connected via Ethernet or one or more other networking protocols such as, for example, Infmiband, Myrinet, or over a memory bus.
  • the servers may run any operating system (OS) (e.g., Windows, Linux).
  • OS operating system
  • each of the servers in a cluster may be responsible for running some portion of an HTM network.
  • the portion of the HTM network dedicated to each server may vary from server to server depending on, for example, the configuration/specification of each server.
  • the CPUs over which an HTM network runs may be located at a single location (e.g., at a datacenter) or at locations remote from one another.
  • an HTM network may be provided as a software platform.
  • the software executables for creating and running the HTM network may be referred to as being part of a "runtime engine.”
  • a runtime engine 172 of an HTM-based system includes, in addition to the executables for running an HTM network 174, a Supervisor entity 176.
  • the Supervisor entity 176 is responsible for, among other things, starting and stopping the HTM network 174 and communicating with external applications (i.e., "tools") 180, 182, 184, each of which are further described below.
  • tools external applications
  • the Supervisor entity 176 may be used to start and stop the HTM network 174, it may not be necessary for the Supervisor entity 176 to be running while the HTM network 174 is in operation.
  • the Supervisor entity 176 uses a description in the net list 178 to configure the HTM network 174.
  • a description in the net list 178 may specify the distribution of nodes across a given set of CPUs.
  • the Supervisor entity 176 may configure an HTM network dynamically if, for example, certain information is not contained in the net list 178.
  • the Supervisor entity 176 may read a net list from a date file.
  • a net list may be specified interactively by a user using one or more tools 180, 182, 184.
  • the Supervisor entity 176 may perform global network actions, distribute nodes across CPUs, and/or coordinate CPU activity/behavior. Further, in one or more embodiments of the present invention, the Supervisor entity 176 may enforce licensing restrictions such as those relating to, for example, the number of usable CPUs, license expiration dates, number of user limitations, and/or the ability to load third-party "plug-ins.”
  • the Supervisor entity 176 may check for software updates on some regular basis. In such embodiments, if there is a software update available, the Supervisor entity 176 may, for example, install the software update and restart the HTM network 174. Further, in one or more embodiments of the present invention, the Supervisor entity 176 may determine and/or select the order in which portions of the HTM network 174 are to be updated.
  • the Supervisor entity 176 may communicate with one or more CPUs (not shown in Figure 15) running the HTM network 174 using, for example, a private or internal application program interface (API). Further, in one or more embodiments of the present invention, the Supervisor entity 176 and the one or more CPUs (not shown in Figure 15) running the HTM network 174 may all be on the same local area network (LAN).
  • LAN local area network
  • the Supervisor entity 176 may run on a CPU separate from one or more CPUs (not shown in Figure 15) running the HTM network 174. However, in one or more other embodiments of the present invention, the Supervisor entity 176 may run on a CPU that runs all or part of the HTM network 174.
  • Figure 18 shows at least a portion of an HTM-based system that runs an HTM network 186 on a single CPU 188.
  • an instance of Supervisor entity 190 may run on CPU 188.
  • a runtime engine 194 may be composed of the software executables for the HTM network 186, the Supervisor entity 190, and the net list 192.
  • FIG. 19 shows at least a portion of an HTM-based system that runs an HTM network 220 on multiple CPUs 222, 224, 226.
  • the CPUs 222, 224, 226 may all be part of the same server (thereby, sharing resources of that server) or they may be distributed over two or more servers.
  • An instance of Supervisor entity 228, along with a net list 230, may run on a separate CPU 232.
  • the Supervisor entity 228 may communicate (across, for example, a switch 234) with instances of "node processing units" (NPUs) 236, 238, 240 running on each of the CPUs 222, 224, 226.
  • NPUs node processing units
  • Each NPU 236, 238, 240 may be a software component that is responsible for running and/or scheduling a portion (i.e., a "sub-net") of the HTM network 220 running on the CPU 222, 224, 226 to which the NPU 236, 238, 240 is respectively allocated.
  • each NPU 236, 238, 240 may receive information from the Supervisor entity 228 describing all or part of the HTM network 220, including information relating to the portion of the HTM network 220 that each NPU 236, 238, 240 will manage. Further, each NPU 236, 238, 240 may be responsible for allocating the memory needed for the nodes, links, and other data structures for the portion of the HTM network 220 for which it is responsible. Further, each NPU 236, 238, 240 may run and/or schedule a portion of the HTM network 220 in some timing relation to at least one other NPU 236, 238, 240.
  • a local net list may be used by an NPU to determine when to update one or more nodes, where "updating" a node may include executing an operation of the node and then updating the state of the node.
  • An NPU may perform such updating based on, for example, one or more timestamps of previous updates of one or more nodes, one or more values (e.g., beliefs) of one or more nodes, priorities of one or more nodes, and/or a set of rules for updating nodes.
  • a runtime engine 242 may be composed of the software executables for the HTM network 220, the Supervisor entity 228, the net list 230, and the NPUs 236, 238, 240.
  • a file server (not shown) may be present to store file information for one or more of the various components shown in Figure 19.
  • there is one NPU per CPU running a portion of an HTM network there is one NPU per CPU running a portion of an HTM network.
  • a runtime engine 1720 running HTM network 174 may interface with one or more tools 180, 182, 184.
  • Each of these tools 180, 182, 184 may be used by a user (e.g., a software developer) to, for example, modify, improve, augment, restrict, configure, or otherwise affect an operation or configuration of the HTM network 174 or a CPU on which the HTM network 174 runs.
  • Configurator tool 180 may be used to create and/or configure an HTM network
  • Trainer tool 182 may be used to create a trained HTM network for a particular application
  • Debugger tool 184 may be used to debug the operation of an HTM network.
  • tools may be provided to, for example, monitor/report performance of an HTM network and/or deploy a designed, trained, and/or debugged HTM network as a running application.
  • one or more embodiments of the present invention may use any number and/or types of different tools to interface with an HTM network.
  • a Supervisor entity e.g., a supervisor entity
  • Supervisor API may support Unicode and/or multi-byte character sets.
  • a Supervisor API may be accessible through, for example, a firewall.
  • One protocol that may be used to facilitate such accessibility involves encoding messages in Extensible Markup Language (XML) and passing them over the Internet (i.e., HTTP transmission). If security is desired or required, then messages may be passed over a secure Internet protocol (e.g., HTTPS transmission).
  • XML Extensible Markup Language
  • HTTPS transmission a secure Internet protocol
  • a Supervisor entity e.g., 176 in Figure 15, 190 in Figure 16, 228 in Figure 17
  • developer/client tools e.g., 180, 182, 184 in Figure 15
  • messages may be passed using means such as, for example, socket connections and/or pipes.
  • a Supervisor API may interact with developer/client tools.
  • the Supervisor API may be used to authenticate one or more client applications attempting to communicate with a Supervisor entity (e.g., 176 in Figure 15, 190 in Figure 16, 228 in Figure 17). If the client is authenticated, the Supervisor API may return session information to the client and connect the client with the Supervisor entity. The Supervisor API may also disconnect the client from the Supervisor entity.
  • a Supervisor entity e.g., 176 in Figure 15, 190 in Figure 16, 228 in Figure 17.
  • a net list describing all or part of an HTM network may be passed from a client to a Supervisor entity through a Supervisor API.
  • a Supervisor API may be used to return state information to the client. State information may include, for example, the beliefs at one or more nodes of the HTM network, whether the HTM network is running, paused, or restarting, the number of nodes in all or part of the HTM network, and the number of CPUs actively running portions of the HTM network.
  • a Supervisor API may be accessed to start, pause and restart, or stop an HTM network.
  • a Supervisor API may be accessed to: return a list of network files that have been stored by a system (e.g., a cluster of servers) used to run an HTM network; load an HTM network from a network file stored locally in a system (e.g., a cluster of servers) usable to run an HTM network; locally save a state of an HTM network in a system (e.g., a cluster of servers) running the HTM network; move one or more nodes from running on one CPU to running on another CPU; turn a debugging feature "on” or "off; retrieve detailed state information of a component in an HTM network; set a state of a component in an HTM network; instruct an HTM network to pause operations after a specific triggering event, where the triggering event may be completion of one complete iteration of the HTM network, completion of updating a given list of nodes, completion of updating one node on each CPU, reaching a particular time, reaching
  • a batch command system may be used to execute one or more operations of a Supervisor API in a particular sequence. Further, in one or more embodiments of the present invention, a batch command system may be used to execute one or more of the same commands on more than one node. Further, in one or more embodiments of the present invention, a batch command system may include the capabilities of a full scripting language (e.g., Python, Perl) so that, for example, 'if statements and loops may be performed easily.
  • Python Python
  • script complex commands e.g., commands: train level 1 of hierarchy until states of level 1 nodes reach a given condition; then turn "off learning in level 1 and train level 2 of hierarchy until states of level 2 nodes reach a given condition, etc.).
  • the Supervisor API may be arranged to handle a failure of any of the hardware components needed to run a particular HTM network. Further, in one or more embodiments of the present invention, the Supervisor API may handle a software failure (e.g., failure of an NPU instance). Further, in one or more embodiments of the present invention, the Supervisor API may handle a communication establishment error. Further, in one or more embodiments of the present invention, the Supervisor API may handle one or more errors in reading a provided net list describing a particular HTM network.
  • an HTM-based system may also have a Node
  • the Node Plug-in API 250 may be used to create new node types.
  • the Node Plug-in API 250 may be used to interface new hardware for running the HTM network 186 and/or implement, for example, new learning algorithms.
  • one or more "plug- ins" may be dynamically loaded when the HTM network 186 is initialized or rebooted. In such a manner, a functionality of a runtime engine running the HTM network 186 may be extended as further described below.

Abstract

L'invention concerne un nœud d'apprentissage spatiotemporel qui est un type de nœud HTM qui apprend des groupes à la fois spatiaux et temporels de schémas d'entrée détectés au cours du temps. Les nœuds d'apprentissage spatiotemporel comprennent des regroupeurs spatiaux qui sont utilisés pour déterminer les groupes spatiaux dans un ensemble de schémas d'entrée détectés. Les nœuds d'apprentissage spatiotemporel comprennent en outre des regroupeurs temporels qui sont utilisés pour déterminer les groupes de schémas d'entrée détectés qui se coproduisent temporellement. Un réseau d'apprentissage spatiotemporel est un réseau hiérarchique qui comprend une pluralité de nœuds d'apprentissage spatiotemporel.
PCT/US2008/055352 2007-02-28 2008-02-28 Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques WO2008106615A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
US90463407P 2007-02-28 2007-02-28
US60/904,634 2007-02-28

Publications (1)

Publication Number Publication Date
WO2008106615A1 true WO2008106615A1 (fr) 2008-09-04

Family

ID=39717043

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2008/055352 WO2008106615A1 (fr) 2007-02-28 2008-02-28 Algorithmes d'apprentissage spatiotemporel dans des réseaux temporels hiérarchiques

Country Status (2)

Country Link
US (2) US8037010B2 (fr)
WO (1) WO2008106615A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11793843B2 (en) 2019-01-10 2023-10-24 Janssen Biotech, Inc. Prostate neoantigens and their uses

Families Citing this family (58)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20060184462A1 (en) * 2004-12-10 2006-08-17 Hawkins Jeffrey C Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
US7739208B2 (en) 2005-06-06 2010-06-15 Numenta, Inc. Trainable hierarchical memory system and method
US20070192267A1 (en) 2006-02-10 2007-08-16 Numenta, Inc. Architecture of a hierarchical temporal memory based system
US8732098B2 (en) 2006-02-10 2014-05-20 Numenta, Inc. Hierarchical temporal memory (HTM) system deployed as web service
US8112367B2 (en) * 2007-02-28 2012-02-07 Numenta, Inc. Episodic memory with a hierarchical temporal memory based system
US8037010B2 (en) 2007-02-28 2011-10-11 Numenta, Inc. Spatio-temporal learning algorithms in hierarchical temporal networks
US8219507B2 (en) * 2007-06-29 2012-07-10 Numenta, Inc. Hierarchical temporal memory system with enhanced inference capability
US8200025B2 (en) * 2007-12-07 2012-06-12 University Of Ottawa Image classification and search
US8175985B2 (en) * 2008-03-19 2012-05-08 Numenta, Inc. Plugin infrastructure for hierarchical temporal memory (HTM) system
US7983998B2 (en) * 2008-03-21 2011-07-19 Numenta, Inc. Feedback in group based hierarchical temporal memory system
US8407166B2 (en) * 2008-06-12 2013-03-26 Numenta, Inc. Hierarchical temporal memory system with higher-order temporal pooling capability
US8195582B2 (en) * 2009-01-16 2012-06-05 Numenta, Inc. Supervision based grouping of patterns in hierarchical temporal memory (HTM)
US8566258B2 (en) * 2009-07-10 2013-10-22 Sony Corporation Markovian-sequence generator and new methods of generating Markovian sequences
EP2548096B1 (fr) 2010-03-15 2017-10-25 Numenta, Inc. Mémoire temporelle utilisant une représentation distribuée sporadique
US11651277B2 (en) 2010-03-15 2023-05-16 Numenta, Inc. Sparse distributed representation for networked processing in predictive system
US8583586B2 (en) * 2011-01-21 2013-11-12 International Business Machines Corporation Mining temporal patterns in longitudinal event data using discrete event matrices and sparse coding
US9053428B2 (en) 2011-07-21 2015-06-09 Qualcomm Incorporated Method and apparatus of robust neural temporal coding, learning and cell recruitments for memory using oscillation
US9147155B2 (en) * 2011-08-16 2015-09-29 Qualcomm Incorporated Method and apparatus for neural temporal coding, learning and recognition
US8645291B2 (en) 2011-08-25 2014-02-04 Numenta, Inc. Encoding of data for processing in a spatial and temporal memory system
US8825565B2 (en) 2011-08-25 2014-09-02 Numenta, Inc. Assessing performance in a spatial and temporal memory system
US8504570B2 (en) 2011-08-25 2013-08-06 Numenta, Inc. Automated search for detecting patterns and sequences in data using a spatial and temporal memory system
US9092735B2 (en) 2011-09-21 2015-07-28 Qualcomm Incorporated Method and apparatus for structural delay plasticity in spiking neural networks
US9111224B2 (en) 2011-10-19 2015-08-18 Qualcomm Incorporated Method and apparatus for neural learning of natural multi-spike trains in spiking neural networks
US10169339B2 (en) 2011-10-31 2019-01-01 Elwha Llc Context-sensitive query enrichment
US8909575B2 (en) 2012-02-29 2014-12-09 Qualcomm Incorporated Method and apparatus for modeling neural resource based synaptic placticity
PT2639749T (pt) 2012-03-15 2017-01-18 Cortical Io Gmbh Métodos, aparelhos e produtos para processamento semântico de texto
US9336774B1 (en) 2012-04-20 2016-05-10 Google Inc. Pattern recognizing engine
US8886574B2 (en) * 2012-06-12 2014-11-11 Siemens Aktiengesellschaft Generalized pattern recognition for fault diagnosis in machine condition monitoring
US9064215B2 (en) 2012-06-14 2015-06-23 Qualcomm Incorporated Learning spike timing precision
US9412067B2 (en) 2012-09-05 2016-08-09 Numenta, Inc. Anomaly detection in spatial and temporal memory system
US9159021B2 (en) 2012-10-23 2015-10-13 Numenta, Inc. Performing multistep prediction using spatial and temporal memory system
US9904889B2 (en) 2012-12-05 2018-02-27 Applied Brain Research Inc. Methods and systems for artificial cognition
US10346739B1 (en) 2013-03-13 2019-07-09 Hrl Laboratories, Llc Method and apparatus for learning, prediction, and recall of spatiotemporal patterns
US10318878B2 (en) 2014-03-19 2019-06-11 Numenta, Inc. Temporal processing scheme and sensorimotor information processing
US10242313B2 (en) * 2014-07-18 2019-03-26 James LaRue Joint proximity association template for neural networks
US10885089B2 (en) 2015-08-21 2021-01-05 Cortical.Io Ag Methods and systems for identifying a level of similarity between a filtering criterion and a data item within a set of streamed documents
US10394851B2 (en) 2014-08-07 2019-08-27 Cortical.Io Ag Methods and systems for mapping data items to sparse distributed representations
US10275705B2 (en) * 2014-08-08 2019-04-30 Vicarious Fpc, Inc. Systems and methods for generating data explanations for neural networks and related systems
WO2016132152A1 (fr) * 2015-02-19 2016-08-25 Magic Pony Technology Limited Interpolation de données visuelles
GB201603144D0 (en) * 2016-02-23 2016-04-06 Magic Pony Technology Ltd Training end-to-end video processes
EP3278559B1 (fr) * 2015-03-31 2021-05-05 Magic Pony Technology Limited Apprentissage de processus vidéo bout-à-bout
GB201604672D0 (en) 2016-03-18 2016-05-04 Magic Pony Technology Ltd Generative methods of super resolution
US9886193B2 (en) 2015-05-15 2018-02-06 International Business Machines Corporation Architecture and implementation of cortical system, and fabricating an architecture using 3D wafer scale integration
US9744668B1 (en) 2015-08-21 2017-08-29 X Development Llc Spatiotemporal robot reservation systems and method
WO2017178808A1 (fr) 2016-04-12 2017-10-19 Magic Pony Technology Limited Traitement de données visuelles à l'aide des réseaux d'énergie
GB201607994D0 (en) 2016-05-06 2016-06-22 Magic Pony Technology Ltd Encoder pre-analyser
US10572221B2 (en) 2016-10-20 2020-02-25 Cortical.Io Ag Methods and systems for identifying a level of similarity between a plurality of data representations
RU2674326C2 (ru) * 2017-02-20 2018-12-06 Автономная некоммерческая образовательная организация высшего образования "Сколковский институт науки и технологий" Способ формирования архитектуры нейросети для классификации объекта, заданного в виде облака точек, способ ее применения для обучения нейросети и поиска семантически схожих облаков точек
US10242087B2 (en) 2017-05-12 2019-03-26 International Business Machines Corporation Cluster evaluation in unsupervised learning of continuous data
US10726072B2 (en) 2017-11-15 2020-07-28 Sap Se Internet of things search and discovery graph engine construction
US10642908B2 (en) * 2017-11-15 2020-05-05 Sap Se Internet of things search and discovery dynamic alteration of results
EP3847803A4 (fr) 2018-09-05 2022-06-15 Vicarious FPC, Inc. Procédé et système de compréhension de concept de machine
US11657291B2 (en) 2019-10-04 2023-05-23 Waymo Llc Spatio-temporal embeddings
US11074163B2 (en) * 2019-11-05 2021-07-27 Jpmorgan Chase Bank, N.A. Method and system for generating unit tests using machine learning
US11681922B2 (en) 2019-11-26 2023-06-20 Numenta, Inc. Performing inference and training using sparse neural network
CN111325165B (zh) * 2020-02-26 2023-05-05 中南大学 考虑空间关系信息的城市遥感影像场景分类方法
US11734332B2 (en) 2020-11-19 2023-08-22 Cortical.Io Ag Methods and systems for reuse of data item fingerprints in generation of semantic maps
CN114387030B (zh) * 2022-01-13 2024-03-15 瑞祥全球购超市有限公司 一种面向网络购物平台的在线用户量的智能分析方法

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6122014A (en) * 1998-09-17 2000-09-19 Motorola, Inc. Modified chroma keyed technique for simple shape coding for digital video
US20050190990A1 (en) * 2004-01-27 2005-09-01 Burt Peter J. Method and apparatus for combining a plurality of images

Family Cites Families (73)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4766534A (en) 1986-10-16 1988-08-23 American Telephone And Telegraph Company, At&T Bell Laboratories Parallel processing network and method
US4845744A (en) 1986-10-16 1989-07-04 American Telephone And Telegraph Company, At&T Bell Laboratories Method of overlaying virtual tree networks onto a message passing parallel processing network
US5255348A (en) 1991-06-14 1993-10-19 Nenov Valeriy I Neural network for learning, recognition and recall of pattern sequences
US7481453B2 (en) 1991-07-09 2009-01-27 Automotive Technologies International, Inc. Inflator system
US6400996B1 (en) 1999-02-01 2002-06-04 Steven M. Hoffberg Adaptive pattern recognition based control system and method
US5729661A (en) 1992-11-24 1998-03-17 Pavilion Technologies, Inc. Method and apparatus for preprocessing input data to a neural network
US7251637B1 (en) 1993-09-20 2007-07-31 Fair Isaac Corporation Context vector generation and retrieval
JPH0877010A (ja) 1994-09-07 1996-03-22 Hitachi Ltd データ分析方法および装置
US7786864B1 (en) 2000-09-08 2010-08-31 Automotive Technologies International, Inc. Vehicular RFID and sensor assemblies
US5712953A (en) 1995-06-28 1998-01-27 Electronic Data Systems Corporation System and method for classification of audio or audio/video signals based on musical content
EP0931388B1 (fr) * 1996-08-29 2003-11-05 Cisco Technology, Inc. Traitement spatio-temporel pour telecommunications
US6028608A (en) 1997-05-09 2000-02-22 Jenkins; Barry System and method of perception-based image generation and encoding
US6516416B2 (en) 1997-06-11 2003-02-04 Prism Resources Subscription access system for use with an untrusted network
US6195622B1 (en) 1998-01-15 2001-02-27 Microsoft Corporation Methods and apparatus for building attribute transition probability models for use in pre-fetching resources
US6636631B2 (en) 1998-06-04 2003-10-21 Matsushita Electric Industrial Co., Ltd. Optical character reading method and system for a document with ruled lines and its application
US6567814B1 (en) 1998-08-26 2003-05-20 Thinkanalytics Ltd Method and apparatus for knowledge discovery in databases
US6751343B1 (en) 1999-09-20 2004-06-15 Ut-Battelle, Llc Method for indexing and retrieving manufacturing-specific digital imagery based on image content
US6468069B2 (en) 1999-10-25 2002-10-22 Jerome H. Lemelson Automatically optimized combustion control
US6625585B1 (en) 2000-02-18 2003-09-23 Bioreason, Inc. Method and system for artificial intelligence directed lead discovery though multi-domain agglomerative clustering
DE60113073T2 (de) 2000-03-10 2006-08-31 Smiths Detection Inc., Pasadena Steuerung für einen industriellen prozes mit einer oder mehreren multidimensionalen variablen
US20020006121A1 (en) 2000-04-27 2002-01-17 Dileep George Adaptive diversity combining for wide band code division multiple access (W-CDMA) based on iterative channel estimation
US7054850B2 (en) 2000-06-16 2006-05-30 Canon Kabushiki Kaisha Apparatus and method for detecting or recognizing pattern by employing a plurality of feature detecting elements
DE60125536T2 (de) 2000-06-30 2007-10-04 British Telecommunications P.L.C. Anordnung zur generierung von elementensequenzen
US6714941B1 (en) 2000-07-19 2004-03-30 University Of Southern California Learning data prototypes for information extraction
AU2001295591A1 (en) 2000-10-13 2002-04-22 Fraunhofer-Gesellschaft Zur Forderung Der Angewandten Forschung E.V. A method for supervised teaching of a recurrent artificial neural network
US6574537B2 (en) 2001-02-05 2003-06-03 The Boeing Company Diagnostic system and method
US6947378B2 (en) 2001-02-28 2005-09-20 Mitsubishi Electric Research Labs, Inc. Dynamic network resource allocation using multimedia content features and traffic features
US6615211B2 (en) 2001-03-19 2003-09-02 International Business Machines Corporation System and methods for using continuous optimization for ordering categorical data sets
US7962482B2 (en) 2001-05-16 2011-06-14 Pandora Media, Inc. Methods and systems for utilizing contextual feedback to generate and modify playlists
US7797062B2 (en) 2001-08-10 2010-09-14 Rockwell Automation Technologies, Inc. System and method for dynamic multi-objective optimization of machine selection, integration and utilization
US20040142325A1 (en) 2001-09-14 2004-07-22 Liat Mintz Methods and systems for annotating biomolecular sequences
US7233781B2 (en) 2001-10-10 2007-06-19 Ochoa Optics Llc System and method for emergency notification content delivery
US6957241B2 (en) 2002-02-14 2005-10-18 Gallitzin Allegheny Llc FFT and FHT engine
US7035431B2 (en) * 2002-02-22 2006-04-25 Microsoft Corporation System and method for probabilistic exemplar-based pattern tracking
US7203635B2 (en) 2002-06-27 2007-04-10 Microsoft Corporation Layered models for context awareness
KR20050097936A (ko) 2003-01-10 2005-10-10 코닌클리케 필립스 일렉트로닉스 엔.브이. 효과적인 예측 이미지 파라미터 추정
US20040148520A1 (en) 2003-01-29 2004-07-29 Rajesh Talpade Mitigating denial of service attacks
US7627171B2 (en) * 2003-07-03 2009-12-01 Videoiq, Inc. Methods and systems for detecting objects of interest in spatio-temporal signals
US8510083B2 (en) 2003-07-31 2013-08-13 The Boeing Company Method, apparatus and computer program product for constructing a diagnostic network model
JP3987013B2 (ja) 2003-09-01 2007-10-03 本田技研工業株式会社 車両周辺監視装置
US7565451B2 (en) 2004-01-23 2009-07-21 Microsoft Corporation Adaptive dispatch of received messages to code using inter-positioned message modification
US7853456B2 (en) 2004-03-05 2010-12-14 Health Outcomes Sciences, Llc Systems and methods for risk stratification of patient populations
US7788109B2 (en) 2004-04-03 2010-08-31 Altusys Corp. Method and apparatus for context-sensitive event correlation with external control in situation-based management
US7565194B2 (en) 2004-05-12 2009-07-21 Zoll Medical Corporation ECG rhythm advisory method
US20060184462A1 (en) 2004-12-10 2006-08-17 Hawkins Jeffrey C Methods, architecture, and apparatus for implementing machine intelligence and hierarchical memory systems
TWI273409B (en) 2005-01-18 2007-02-11 Sunplus Technology Co Ltd Pre-fetch technology by read merge mechanism
JP2006285899A (ja) 2005-04-05 2006-10-19 Sony Corp 学習装置および学習方法、生成装置および生成方法、並びにプログラム
US7577651B2 (en) 2005-04-28 2009-08-18 Yahoo! Inc. System and method for providing temporal search results in response to a search query
US7809722B2 (en) 2005-05-09 2010-10-05 Like.Com System and method for enabling search and retrieval from image files based on recognized information
US7739208B2 (en) 2005-06-06 2010-06-15 Numenta, Inc. Trainable hierarchical memory system and method
WO2007087594A2 (fr) 2006-01-25 2007-08-02 Espeed, Inc. Systèmes et procédés permettant de faciliter l'exécution de conventions de rachat
US20070192267A1 (en) * 2006-02-10 2007-08-16 Numenta, Inc. Architecture of a hierarchical temporal memory based system
US20080208966A1 (en) 2007-02-28 2008-08-28 Numenta, Inc. Hierarchical Temporal Memory (HTM) System Deployed as Web Service
US8732098B2 (en) 2006-02-10 2014-05-20 Numenta, Inc. Hierarchical temporal memory (HTM) system deployed as web service
US7941389B2 (en) 2006-02-10 2011-05-10 Numenta, Inc. Hierarchical temporal memory based system including nodes with input or output variables of disparate properties
WO2007095585A2 (fr) 2006-02-14 2007-08-23 Edsa Micro Corporation Systemes et procedes pour la surveillance du systeme en temps reel et l'analyse predictive
US7840395B2 (en) 2006-03-10 2010-11-23 Edsa Micro Corporation Systems and methods for predictive monitoring including real-time strength and security analysis in an electrical power distribution system
US7840396B2 (en) 2006-03-10 2010-11-23 Edsa Micro Corporation Systems and methods for determining protective device clearing times used for providing real-time predictions about arc flash events
CA2646423A1 (fr) 2006-03-10 2007-09-20 Edsa Micro Corporation Systemes et procedes d'evaluation de dispositifs de protection en temps reel dans un systeme de distribution d'energie electrique
US7844440B2 (en) 2006-07-07 2010-11-30 Edsa Micro Corporation Systems and methods for real-time dynamic simulation of uninterruptible power supply solutions and their control logic systems
WO2008067326A2 (fr) 2006-11-28 2008-06-05 Numenta, Inc. Mise en commun temporelle basée sur des groupes
US8037010B2 (en) 2007-02-28 2011-10-11 Numenta, Inc. Spatio-temporal learning algorithms in hierarchical temporal networks
US7941392B2 (en) 2007-02-28 2011-05-10 Numenta, Inc. Scheduling system and method in a hierarchical temporal memory based system
US8112367B2 (en) 2007-02-28 2012-02-07 Numenta, Inc. Episodic memory with a hierarchical temporal memory based system
US8219507B2 (en) 2007-06-29 2012-07-10 Numenta, Inc. Hierarchical temporal memory system with enhanced inference capability
US7676458B2 (en) 2007-08-28 2010-03-09 International Business Machines Corporation System and method for historical diagnosis of sensor networks
WO2009052407A1 (fr) 2007-10-18 2009-04-23 Numenta, Inc. Système et procédé pour une détermination de topologie automatique dans un réseau temporel hiérarchique
US8175984B2 (en) 2007-12-05 2012-05-08 Numenta, Inc. Action based learning
US8175985B2 (en) 2008-03-19 2012-05-08 Numenta, Inc. Plugin infrastructure for hierarchical temporal memory (HTM) system
US7983998B2 (en) 2008-03-21 2011-07-19 Numenta, Inc. Feedback in group based hierarchical temporal memory system
US8407166B2 (en) 2008-06-12 2013-03-26 Numenta, Inc. Hierarchical temporal memory system with higher-order temporal pooling capability
US8195582B2 (en) 2009-01-16 2012-06-05 Numenta, Inc. Supervision based grouping of patterns in hierarchical temporal memory (HTM)
EP2548096B1 (fr) 2010-03-15 2017-10-25 Numenta, Inc. Mémoire temporelle utilisant une représentation distribuée sporadique

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6122014A (en) * 1998-09-17 2000-09-19 Motorola, Inc. Modified chroma keyed technique for simple shape coding for digital video
US20050190990A1 (en) * 2004-01-27 2005-09-01 Burt Peter J. Method and apparatus for combining a plurality of images

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
HAWKINS J. AND GEORGE D.: "Hierarchical Temporal Meory - Concepts, Theory, and Terminology", 10 May 2006 (2006-05-10), Retrieved from the Internet <URL:http://www.neurosecurity.com/whitepapers/Numenta_HTM_Concepts.pdf> *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US11793843B2 (en) 2019-01-10 2023-10-24 Janssen Biotech, Inc. Prostate neoantigens and their uses

Also Published As

Publication number Publication date
US20080208783A1 (en) 2008-08-28
US20120005134A1 (en) 2012-01-05
US8504494B2 (en) 2013-08-06
US8037010B2 (en) 2011-10-11

Similar Documents

Publication Publication Date Title
US8037010B2 (en) Spatio-temporal learning algorithms in hierarchical temporal networks
US7937342B2 (en) Method and apparatus for detecting spatial patterns
US8219507B2 (en) Hierarchical temporal memory system with enhanced inference capability
US8175984B2 (en) Action based learning
US8121961B2 (en) Feedback in group based hierarchical temporal memory system
US11610131B2 (en) Ensembling of neural network models
US7624085B2 (en) Hierarchical based system for identifying object using spatial and temporal patterns
US8195582B2 (en) Supervision based grouping of patterns in hierarchical temporal memory (HTM)
US8112367B2 (en) Episodic memory with a hierarchical temporal memory based system
De Paula et al. Building Up Conceptual Spaces: An ESOM Supported Strategy

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 08743605

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 08743605

Country of ref document: EP

Kind code of ref document: A1